1. Add eval script

2. Add inference script
3. Add load model script
4. Add some functions into ctc_reader

fluid/ocr_recognition/crnn_ctc_model.py

-import paddle.v2.fluid as fluid
+import paddle.fluid as fluid
 
 
 def conv_bn_pool(input,
@@ -7,7 +7,8 @@ def conv_bn_pool(input,
                  act="relu",
                  param=None,
                  bias=None,
-                 param_0=None):
+                 param_0=None,
+                 is_test=False):
     tmp = input
     for i in xrange(group):
         tmp = fluid.layers.conv2d(
@@ -19,14 +20,23 @@ def conv_bn_pool(input,
             act=None,  # LinearActivation
             use_cudnn=True)
         tmp = fluid.layers.batch_norm(
-            input=tmp, act=act, param_attr=param, bias_attr=bias)
+            input=tmp,
+            act=act,
+            param_attr=param,
+            bias_attr=bias,
+            is_test=is_test)
     tmp = fluid.layers.pool2d(
         input=tmp, pool_size=2, pool_type='max', pool_stride=2, use_cudnn=True)
 
     return tmp
 
 
-def ocr_convs(input, num, with_bn, regularizer=None, gradient_clip=None):
+def ocr_convs(input,
+              num,
+              with_bn,
+              regularizer=None,
+              gradient_clip=None,
+              is_test=False):
     assert (num % 4 == 0)
 
     b = fluid.ParamAttr(
@@ -42,10 +52,11 @@ def ocr_convs(input, num, with_bn, regularizer=None, gradient_clip=None):
         gradient_clip=gradient_clip,
         initializer=fluid.initializer.Normal(0.0, 0.01))
     tmp = input
-    tmp = conv_bn_pool(tmp, 2, [16, 16], param=w1, bias=b, param_0=w0)
-    tmp = conv_bn_pool(tmp, 2, [32, 32], param=w1, bias=b)
-    tmp = conv_bn_pool(tmp, 2, [64, 64], param=w1, bias=b)
-    tmp = conv_bn_pool(tmp, 2, [128, 128], param=w1, bias=b)
+    tmp = conv_bn_pool(
+        tmp, 2, [16, 16], param=w1, bias=b, param_0=w0, is_test=is_test)
+    tmp = conv_bn_pool(tmp, 2, [32, 32], param=w1, bias=b, is_test=is_test)
+    tmp = conv_bn_pool(tmp, 2, [64, 64], param=w1, bias=b, is_test=is_test)
+    tmp = conv_bn_pool(tmp, 2, [128, 128], param=w1, bias=b, is_test=is_test)
     return tmp
 
 
@@ -53,9 +64,15 @@ def encoder_net(images,
                 num_classes,
                 rnn_hidden_size=200,
                 regularizer=None,
-                gradient_clip=None):
+                gradient_clip=None,
+                is_test=False):
     conv_features = ocr_convs(
-        images, 8, True, regularizer=regularizer, gradient_clip=gradient_clip)
+        images,
+        8,
+        True,
+        regularizer=regularizer,
+        gradient_clip=gradient_clip,
+        is_test=is_test)
     sliced_feature = fluid.layers.im2sequence(
         input=conv_features,
         stride=[1, 1],
@@ -111,12 +128,12 @@ def encoder_net(images,
                              size=num_classes + 1,
                              param_attr=w_attr,
                              bias_attr=b_attr)
+
     return fc_out
 
 
 def ctc_train_net(images, label, args, num_classes):
     regularizer = fluid.regularizer.L2Decay(args.l2)
-    # gradient_clip=fluid.clip.GradientClipByValue(args.max_clip, args.min_clip)
     gradient_clip = None
     fc_out = encoder_net(
         images,
@@ -127,7 +144,7 @@ def ctc_train_net(images, label, args, num_classes):
     cost = fluid.layers.warpctc(
         input=fc_out,
         label=label,
-        size=num_classes + 1,
+        #        size=num_classes + 1,
         blank=num_classes,
         norm_by_times=True)
     avg_cost = fluid.layers.mean(x=cost)
@@ -142,3 +159,27 @@ def ctc_train_net(images, label, args, num_classes):
     error_evaluator = fluid.evaluator.EditDistance(
         input=decoded_out, label=casted_label)
     return avg_cost, error_evaluator
+
+
+def ctc_infer(images, num_classes):
+    fc_out = encoder_net(images, num_classes, is_test=True)
+    return fluid.layers.ctc_greedy_decoder(input=fc_out, blank=num_classes)
+
+
+def ctc_eval(images, label, num_classes):
+    fc_out = encoder_net(images, num_classes, is_test=True)
+    decoded_out = fluid.layers.ctc_greedy_decoder(
+        input=fc_out, blank=num_classes)
+
+    casted_label = fluid.layers.cast(x=label, dtype='int64')
+    error_evaluator = fluid.evaluator.EditDistance(
+        input=decoded_out, label=casted_label)
+
+    cost = fluid.layers.warpctc(
+        input=fc_out,
+        label=label,
+        #size=num_classes + 1,
+        blank=num_classes,
+        norm_by_times=True)
+
+    return error_evaluator, cost

fluid/ocr_recognition/ctc_reader.py

@@ -4,6 +4,10 @@ import numpy as np
 from PIL import Image
 
 from paddle.v2.image import load_image
+import paddle.v2 as paddle
+
+NUM_CLASSES = 10784
+DATA_SHAPE = [1, 48, 512]
 
 
 class DataGenerator(object):
@@ -15,10 +19,10 @@ class DataGenerator(object):
         Reader interface for training.
 
         :param img_root_dir: The root path of the image for training.
-        :type file_list: str 
+        :type file_list: str
 
         :param img_label_list: The path of the <image_name, label> file for training.
-        :type file_list: str 
+        :type file_list: str
 
         '''
 
@@ -76,7 +80,7 @@ class DataGenerator(object):
         Reader interface for inference.
 
         :param img_root_dir: The root path of the images for training.
-        :type file_list: str 
+        :type file_list: str
 
         :param img_label_list: The path of the <image_name, label> file for testing.
         :type file_list: list
@@ -95,3 +99,28 @@ class DataGenerator(object):
                 yield img, label
 
         return reader
+
+
+def num_classes():
+    return NUM_CLASSES
+
+
+def data_shape():
+    return DATA_SHAPE
+
+
+def train(batch_size):
+    generator = DataGenerator()
+    return generator.train_reader(
+        "/home/disk1/wanghaoshuang/models/fluid/ocr_recognition/data/train_images/",
+        "/home/disk1/wanghaoshuang/models/fluid/ocr_recognition/data/train.list",
+        batch_size)
+
+
+def test(batch_size=1):
+    generator = DataGenerator()
+    return paddle.batch(
+        generator.test_reader(
+            "/home/disk1/wanghaoshuang/models/fluid/ocr_recognition/data/test_images/",
+            "/home/disk1/wanghaoshuang/models/fluid/ocr_recognition/data/test.list"
+        ), batch_size)

fluid/ocr_recognition/ctc_train.py

 """Trainer for OCR CTC model."""
 import paddle.v2 as paddle
-import paddle.v2.fluid as fluid
+import paddle.fluid as fluid
 import dummy_reader
+import ctc_reader
 import argparse
+from load_model import load_param
 import functools
 import sys
 from utility import add_arguments, print_arguments, to_lodtensor, get_feeder_data
@@ -12,7 +14,7 @@ parser = argparse.ArgumentParser(description=__doc__)
 add_arg = functools.partial(add_arguments, argparser=parser)
 # yapf: disable
 add_arg('batch_size',     int,   32,     "Minibatch size.")
-add_arg('pass_num',       int,   32,     "# of training epochs.")
+add_arg('pass_num',       int,   100,     "# of training epochs.")
 add_arg('log_period',     int,   1000,   "Log period.")
 add_arg('learning_rate',  float, 1.0e-3, "Learning rate.")
 add_arg('l2',             float, 0.0004, "L2 regularizer.")
@@ -24,6 +26,15 @@ add_arg('device',         int,   0,      "Device id.'-1' means running on CPU"
                                          "while '0' means GPU-0.")
 # yapf: disable
 
+def load_parameter(place):
+    params = load_param('./name.map', './data/model/results_without_avg_window/pass-00000/')
+    for name in params:
+        #        print "param: %s" % name
+        t = fluid.global_scope().find_var(name).get_tensor()
+        t.set(params[name], place)
+
+
+
 def train(args, data_reader=dummy_reader):
     """OCR CTC training"""
     num_classes = data_reader.num_classes()
@@ -42,38 +53,54 @@ def train(args, data_reader=dummy_reader):
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
 
+    load_parameter(place)
+
     inference_program = fluid.io.get_inference_program(error_evaluator)
+
+    # evaluate model on test data
+    error_evaluator.reset(exe)
+    for data in test_reader():
+        exe.run(inference_program, feed=get_feeder_data(data, place))
+    _, test_seq_error = error_evaluator.eval(exe)
+    print "\nEnd pass[%d]; Test seq error: %s.\n" % (
+            -1, str(test_seq_error[0]))
+
+
+
     for pass_id in range(args.pass_num):
         error_evaluator.reset(exe)
-        batch_id = 0
+        batch_id = 1
+        total_loss = 0.0
+        total_seq_error = 0.0
         # train a pass
         for data in train_reader():
-            loss, batch_edit_distance = exe.run(
+            batch_loss, _, batch_seq_error = exe.run(
                 fluid.default_main_program(),
                 feed=get_feeder_data(data, place),
                 fetch_list=[avg_cost] + error_evaluator.metrics)
-            if batch_id % args.log_period == 0:
-                print "Pass[%d]-batch[%d]; Loss: %s; Word error: %s." % (
-                    pass_id, batch_id, loss[0], batch_edit_distance[0] / float(args.batch_size))
+            total_loss += batch_loss[0]
+            total_seq_error += batch_seq_error[0]
+            if batch_id % 10 == 1:
+                print '.',
+                sys.stdout.flush()
+            if batch_id % args.log_period == 1:
+                print "\nPass[%d]-batch[%d]; Avg Warp-CTC loss: %s; Avg seq error: %s." % (
+                    pass_id, batch_id, total_loss / batch_id, total_seq_error / (batch_id * args.batch_size))
                 sys.stdout.flush()
             batch_id += 1
 
-        train_edit_distance = error_evaluator.eval(exe)
-        print "End pass[%d]; Train word error: %s.\n" % (
-            pass_id, str(train_edit_distance[0]))
-
         # evaluate model on test data
         error_evaluator.reset(exe)
         for data in test_reader():
             exe.run(inference_program, feed=get_feeder_data(data, place))
-        test_edit_distance = error_evaluator.eval(exe)
-        print "End pass[%d]; Test word error: %s.\n" % (
-            pass_id, str(test_edit_distance[0]))
+        _, test_seq_error = error_evaluator.eval(exe)
+        print "\nEnd pass[%d]; Test seq error: %s.\n" % (
+            pass_id, str(test_seq_error[0]))
 
 def main():
     args = parser.parse_args()
     print_arguments(args)
-    train(args, data_reader=dummy_reader)
+    train(args, data_reader=ctc_reader)
 
 if __name__ == "__main__":
     main()

fluid/ocr_recognition/dummy_reader.py

@@ -32,7 +32,7 @@ def _read_creater(num_sample=1024, min_seq_len=1, max_seq_len=10):
     return reader
 
 
-def train(batch_size, num_sample=16):
+def train(batch_size, num_sample=128):
     """Get train dataset reader."""
     return paddle.batch(_read_creater(num_sample=num_sample), batch_size)
 

fluid/ocr_recognition/eval.py

+import paddle.v2 as paddle
+import paddle.fluid as fluid
+from load_model import load_param
+from utility import get_feeder_data
+from crnn_ctc_model import ctc_eval
+import ctc_reader
+import dummy_reader
+
+
+def load_parameter(place):
+    params = load_param('./name.map', './data/model/results/pass-00062/')
+    for name in params:
+        print "param: %s" % name
+        t = fluid.global_scope().find_var(name).get_tensor()
+        t.set(params[name], place)
+
+
+def evaluate(eval=ctc_eval, data_reader=dummy_reader):
+    """OCR inference"""
+    num_classes = data_reader.num_classes()
+    data_shape = data_reader.data_shape()
+    # define network
+    images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
+    label = fluid.layers.data(
+        name='label', shape=[1], dtype='int32', lod_level=1)
+    evaluator, cost = eval(images, label, num_classes)
+
+    # data reader
+    test_reader = data_reader.test()
+    # prepare environment
+    place = fluid.CUDAPlace(0)
+    #place = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    exe.run(fluid.default_startup_program())
+    print fluid.default_main_program()
+    load_parameter(place)
+    evaluator.reset(exe)
+    count = 0
+    for data in test_reader():
+        count += 1
+        print 'Process samples: %d\r' % (count, ),
+        result, avg_distance, avg_seq_error = exe.run(
+            fluid.default_main_program(),
+            feed=get_feeder_data(data, place),
+            fetch_list=[cost] + evaluator.metrics)
+    avg_distance, avg_seq_error = evaluator.eval(exe)
+    print "avg_distance: %s; avg_seq_error: %s" % (avg_distance, avg_seq_error)
+
+
+def main():
+    evaluate(data_reader=ctc_reader)
+
+
+if __name__ == "__main__":
+    main()

fluid/ocr_recognition/inference.py

+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+from load_model import load_param
+from utility import get_feeder_data
+from crnn_ctc_model import ctc_infer
+import ctc_reader
+import dummy_reader
+
+
+def load_parameter(place):
+    params = load_param('./name.map', './data/model/results/pass-00062/')
+    for name in params:
+        print "param: %s" % name
+        t = fluid.global_scope().find_var(name).get_tensor()
+        t.set(params[name], place)
+
+
+def inference(infer=ctc_infer, data_reader=dummy_reader):
+    """OCR inference"""
+    num_classes = data_reader.num_classes()
+    data_shape = data_reader.data_shape()
+    # define network
+    images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
+    sequence, tmp = infer(images, num_classes)
+    fluid.layers.Print(tmp)
+    # data reader
+    test_reader = data_reader.test()
+    # prepare environment
+    place = fluid.CUDAPlace(0)
+    exe = fluid.Executor(place)
+    exe.run(fluid.default_startup_program())
+
+    load_parameter(place)
+
+    for data in test_reader():
+        result = exe.run(fluid.default_main_program(),
+                         feed=get_feeder_data(
+                             data, place, need_label=False),
+                         fetch_list=[tmp])
+        print "result: %s" % (list(result[0].flatten()), )
+
+
+def main():
+    inference(data_reader=ctc_reader)
+
+
+if __name__ == "__main__":
+    main()

fluid/ocr_recognition/load_model.py

+import sys
+import numpy as np
+import ast
+
+
+def load_parameter(file_name):
+    with open(file_name, 'rb') as f:
+        f.read(16)  # skip header.
+        return np.fromfile(f, dtype=np.float32)
+
+
+def load_param(name_map_file, old_param_dir):
+    result = {}
+    name_map = {}
+    shape_map = {}
+    with open(name_map_file, 'r') as map_file:
+        for param in map_file:
+            old_name, new_name, shape = param.strip().split('=')
+            name_map[new_name] = old_name
+            shape_map[new_name] = ast.literal_eval(shape)
+
+    for new_name in name_map:
+        result[new_name] = load_parameter("/".join(
+            [old_param_dir, name_map[new_name]])).reshape(shape_map[new_name])
+    return result
+
+
+if __name__ == "__main__":
+    name_map_file = "./name.map"
+    old_param_dir = "./data/model/results/pass-00062/"
+    result = load_param(name_map_file, old_param_dir)
+    for p in result:
+        print "name: %s; param.shape: %s" % (p, result[p].shape)

fluid/ocr_recognition/utility.py

@@ -18,7 +18,7 @@ from __future__ import division
 from __future__ import print_function
 import distutils.util
 import numpy as np
-from paddle.v2.fluid import core
+from paddle.fluid import core
 
 
 def print_arguments(args):
@@ -77,11 +77,14 @@ def to_lodtensor(data, place):
     return res
 
 
-def get_feeder_data(data, place):
+def get_feeder_data(data, place, need_label=True):
     pixel_tensor = core.LoDTensor()
     pixel_data = None
     pixel_data = np.concatenate(
         map(lambda x: x[0][np.newaxis, :], data), axis=0).astype("float32")
     pixel_tensor.set(pixel_data, place)
     label_tensor = to_lodtensor(map(lambda x: x[1], data), place)
-    return {"pixel": pixel_tensor, "label": label_tensor}
+    if need_label:
+        return {"pixel": pixel_tensor, "label": label_tensor}
+    else:
+        return {"pixel": pixel_tensor}